Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FedGS: Federated Gradient Scaling for Heterogeneous Medical Image Segmentation

Philip Schutte, Valentina Corbetta, Regina Beets-Tan, Wilson Silva

TL;DR

This paper tackles privacy-preserving federated segmentation in heterogeneous multi-center data by introducing FedGS, a gradient-scaling aggregation method that emphasizes small, underrepresented lesions without modifying local training. FedGS computes a cumulatively scaled server update that biases aggregation toward challenging samples through η_t, which is derived from per-image difficulty δ_x based on inverse mask area a^{-1}; the method is dataset-adaptive via base log scaling l and threshold τ. Empirical results on PolypGen and LiTS across UTNet and SD-UTNet show that FedGS substantially improves small-lesion Dice (DiceS) while preserving overall Dice and DiceL, with LiTS benefiting more from small-lesion gains and only modest training overhead. The work demonstrates that gradient-level reweighting at aggregation time can mitigate content-based heterogeneity and enhance performance on clinically important yet underrepresented targets, offering a practical, scalable enhancement to federated medical image segmentation.

Abstract

Federated Learning (FL) in Deep Learning (DL)-automated medical image segmentation helps preserving privacy by enabling collaborative model training without sharing patient data. However, FL faces challenges with data heterogeneity among institutions, leading to suboptimal global models. Integrating Disentangled Representation Learning (DRL) in FL can enhance robustness by separating data into distinct representations. Existing DRL methods assume heterogeneity lies solely in style features, overlooking content-based variability like lesion size and shape. We propose FedGS, a novel FL aggregation method, to improve segmentation performance on small, under-represented targets while maintaining overall efficacy. FedGS demonstrates superior performance over FedAvg, particularly for small lesions, across PolypGen and LiTS datasets. The code and pre-trained checkpoints are available at the following link: https://github.com/Trustworthy-AI-UU-NKI/Federated-Learning-Disentanglement

FedGS: Federated Gradient Scaling for Heterogeneous Medical Image Segmentation

TL;DR

This paper tackles privacy-preserving federated segmentation in heterogeneous multi-center data by introducing FedGS, a gradient-scaling aggregation method that emphasizes small, underrepresented lesions without modifying local training. FedGS computes a cumulatively scaled server update that biases aggregation toward challenging samples through η_t, which is derived from per-image difficulty δ_x based on inverse mask area a^{-1}; the method is dataset-adaptive via base log scaling l and threshold τ. Empirical results on PolypGen and LiTS across UTNet and SD-UTNet show that FedGS substantially improves small-lesion Dice (DiceS) while preserving overall Dice and DiceL, with LiTS benefiting more from small-lesion gains and only modest training overhead. The work demonstrates that gradient-level reweighting at aggregation time can mitigate content-based heterogeneity and enhance performance on clinically important yet underrepresented targets, offering a practical, scalable enhancement to federated medical image segmentation.

Abstract

Federated Learning (FL) in Deep Learning (DL)-automated medical image segmentation helps preserving privacy by enabling collaborative model training without sharing patient data. However, FL faces challenges with data heterogeneity among institutions, leading to suboptimal global models. Integrating Disentangled Representation Learning (DRL) in FL can enhance robustness by separating data into distinct representations. Existing DRL methods assume heterogeneity lies solely in style features, overlooking content-based variability like lesion size and shape. We propose FedGS, a novel FL aggregation method, to improve segmentation performance on small, under-represented targets while maintaining overall efficacy. FedGS demonstrates superior performance over FedAvg, particularly for small lesions, across PolypGen and LiTS datasets. The code and pre-trained checkpoints are available at the following link: https://github.com/Trustworthy-AI-UU-NKI/Federated-Learning-Disentanglement
Paper Structure (17 sections, 4 equations, 2 figures, 1 table)

This paper contains 17 sections, 4 equations, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methodology
  3. FedGS overview
  4. Small lesion classification and difficulty estimation
  5. Experiments
  6. Model architectures
  7. UTNet
  8. SD-UTNet
  9. Datasets
  10. PolypGen
  11. Small polyp classification
  12. LiTS
  13. Small tumor classification
  14. Pre-processing
  15. Training setup
  16. ...and 2 more sections

Figures (2)

  • Figure 1: Overview of the FedGS aggregation method. The segmentation masks depicted in the Figure come from the PolypGen dataset.
  • Figure 2: Plot of Equation \ref{['eq:diff_estimation']} that calculates the difficulty factor $\delta_x$ for images with the inverse area $a^{-1}$ as input. Additionally, five images from PolypGen that contain small polyps are shown in the graph. These images vary greatly in their total mask area and thus illustrate how the estimated difficulty changes based on $a^{-1}$.